Hi Olivier,
I understand well your interest and your volunteering however I do not see any practicable way for achieving your objective.
Grid computing - by design - relies on network infrastructure and on the ability to connect to network / internet for downloading work and uploading results.
Even if you would try to "fool" boinc resp. WCG using a connected "shadow" crunching machine and manually distributing the WU on the non-connected machines and later collecting manually the computed work for reporting it over the front-end "shadow" machine, this approach will not be very reliable (think for example about the result validation) and it will be highly time consuming for you regarding the necessary manual work.
Maybe some other members could propose some valuable ideas regarding your question? ...
Thank you for your contribution.
Happy crunching,
Yves

Be aware that copying files by USB or CD/DVD between the internet connected system and the isolated network risks copying virus or other malware.
The network was isolated for reasons such as this.
If you are not the owner you need written permission from them.
With permission you would be better off using a firewall to alllow connection from the isolated network to just WCG.
This could be a dedicated firewall appliance or a system running BSD or Linux with 2 ethernet ports and firewall configured.

Paul.

Actually the only part that needs to be 'portable' is the BOINC data directory and a host that has dual client installation [if the host is meant to compute too when not used as intermediate file transfer agent], 1 client pointing to the portable medium, set to not compute but networking active during the upload/download phase, making sure to suspend BOINC networking as else you risk the connect counter to go off kilter.

Apart from that a little rigor in discipline such as alternating the clients between online and offline, where the offline client machines are 'always' set to offline.

Buffering is probably practical up to 5-6 days computing worth where the shortest standard deadline is 7 days, obviously excluding FAHB/FAH2 which is just 24 hours.

The hassle comes with lots of satisfaction of course if adding 16-32 and more threads being put to volunteer computing use.

On my side, after a rigorous risk assessment, if the risks are acceptable, I would prefer a dedicated firewall-based approach instead of moving data manually using USB device (so called: "Adidas network" i.e. per pedibus).
Manual data moving is both time consuming AND risky.
Cheers,
Yves
---
PS: The term "Adidas network" comes originally from a friend of mine who is pharmaceutical GMP inspector. I like this term and I use it in my trainings and audits